Machine learning helps discover CHRONO as part of the body’s internal clock

Using the above empirical distributions and a modified version of the Naïve Bayes learning algorithm, we quantified the evidence provided by each feature that a given gene is a member of the core circadian network [22],[32]. We relied on the prior assumption, informed by experimenter judgment, that increasing possession of each of these features lends increasing evidence of a role in the circadian clock. We used the empirical cumulative distribution function (ECDF) describing exemplar “clock genes” or “nonclock genes” to estimate the probabilities that a randomly selected “clock gene” or “nonclock gene” would possess a metric value at least as extreme as the one observed. We term the ratio of these probabilities a “circadian evidence factor” (Materials and Methods, Eq. 4). The evidence factors arising from particular features and metric value are shown in the right panels of Figure 1B–E).

The evidence amassed from all five features is encapsulated by a “combined evidence factor.” Computation of combined evidence factors requires knowledge of the joint cumulative probability distributions for these features among both “clock genes” and “nonclock genes.” These joint cumulative distribution functions are “learned” from the examples under the “Naïve” assumption of conditional independence. Evidence factors from each individual feature are multiplied to calculate the combined evidence factor (,). This approach differs from the standard Naïve Bayes statistical learning approach only in that cumulative distribution functions are used rather than probability density functions. Materials and Methods Eq. 6

We ranked genes based on this combined evidence. The top 20 candidates (Figure 1F) include 10 of the exemplar clock components along with Tef[33] and Nfil3[34], two genes with established circadian functions. Moreover, Wee1, a canonical cell cycle gene, is known to be regulated by the circadian clock [35]. Although, to our knowledge, the hypothesis that Wee1 directly regulates clock function has not been tested. Inspecting the top 50 ranked genes, several other genes known to be involved in the circadian clockworks appear. These include Dbp[36], Insig2[37], and Nampt[38].

In order to evaluate the utility of this ranking in the discovery of novel clock genes, we applied 10-fold cross-validation. We sequentially removed all possible pairs of clock components from the exemplar distribution, ignoring our prior knowledge of their role in orchestrating circadian rhythms. In each case, we then recomputed the combined evidence factors based on this reduced knowledgebase and tested our ability to “rediscover” these clock genes using different ranking cutoffs. Based on this analysis, we estimate that ∼50% of true clock components would be recovered by screening the top 50 genes (Figure S1A). We also compared the use of evidence factors with two prepackaged machine learning algorithms. Using the same features, we ranked genes using a Gaussian Naïve Bayes classifier and a Flexible Naïve Bayes classifier [39]. The three methods all yield comparable performances using cutoffs less than ∼1,000, but the evidence factor method outperforms the other two beyond this point. Importantly, the top candidates from all three methods show a very high degree of overlap (Figure S1B).

Only rankings from the evidence factor approach were used in selecting genes for further screening. However, results from all three probabilistic learning methods are presented in the Supporting Information section. The cycling feature makes the largest single contribution to the combined evidence factors, but it does not completely dominate this ranking. Hundreds of genes demonstrate strong cycling in the tissues analyzed and other features determine the relative ranking among these. Moreover, some candidates, like Hdac11, are largely prioritized based on the combined strength of other features.

Given the rarity of bona fide clock genes, any method that is not 100% specific will result in a number of false positives. As the ranking cutoff is increased, the number of nonclock genes incorrectly identified will also increase. As in other screening applications where one is searching for a “needle in a haystack,” a secondary validation of candidates is needed. Assuming different numbers for the total number of core clock components, we estimated the false positive rate for different screening cutoffs (). Figure S1C

The ultimate value of this approach will be determined by its ability to identify previously unrecognized clock components. We tested the top 25 novel candidates for physical interactions with a subset of proteins from the negative arm of the molecular clock (BMAL1, BMAL2, CLOCK, NPAS2, CRY1, CRY2. PER1, PER2, and PER3). Three of these candidates (Gm129, Ifitm1, and Cbs) demonstrated both physical binding with at least one of the included clock components and a statistically significant change in circadian reporter period after knockdown in the NIH 3T3 model system (Figure S2). Of note, although Gm129 might have been identified simply by its strong cycling, Cbs and Ifitm1 are identified by virtue of a combination of features. Bellow we present a more detailed investigation of the previously uncharacterized candidate, Gm129, here renamed Chrono. These data show that Chrono meets the formal definition of a mammalian circadian clock gene.

Our previous microarray data suggested that Chrono expression cycles with a 24-h period in liver, pituitary, and NIH 3T3 cells [25],[40]. We used quantitative PCR (qPCR) to confirm cycling in the liver and further evaluated transcript cycling in skeletal muscle and white fat (Figure 2). The circadian oscillations in Chrono expression are of a similar magnitude to those observed for known clock factors Nr1d1 and Per2. Consistent with our results, temporal profiling in rat skeletal muscle [41] and lung [42], as well as mouse SCN [43], also revealed daily oscillations in Chrono expression. Several genome-wide, ChIP-seq studies in mouse liver [43]–[45] have identified the E-boxes in the Chrono gene promoter among those genomic regions most tightly bound by BMAL1 protein. Time course microarray studies from SCN and liver demonstrate that Chrono expression is reduced in Clock mutant animals and loses circadian rhythmicity (Figure S3A) [46],[47]. Moreover, Chrono expression becomes arrhythmic in the livers of Cry1/Cry2 double knockout animals (Figure S3B) [48]. In total, Chrono demonstrates robust circadian expression in multiple tissues and appears to be directly regulated by the molecular clock.

We employed a mammalian two-hybrid screen to identify physical interactions between CHRONO and a subset of known clock components. As expected, many core clock proteins physically interacted, as indicated by specific activation of a UAS:Luc reporter in transfected Human Embryonic Kidney 293 cells containing the SV40 T-Antigen (HEK 293T) (Figure 3A, Table S1). Interactions between CHRONO and both BMAL1 and PER2 were also observed, with >20-fold induction of luciferase activity. BMAL1–CHRONO and PER2–CHRONO complex formation were confirmed through co-immunoprecipitation (co-IP) (Figure 3B and C). Bi-molecular Fluorescence Complementation (BiFC) using Venus, an enhanced yellow fluorescent protein (YFP), was then used to map BMAL1/CHRONO interactions to cell nuclei (Figure 3D). Notably, when S-tagged CHRONO was overexpressed with both BMAL1 and CLOCK BiFC fusion proteins, CHRONO appeared to colocalize with the CLOCK/BMAL1 heterodimer in nuclear bodies, suggesting that CHRONO continues to interact with BMAL1 while part of this functional circadian complex.

To evaluate the functional consequences of these physical interactions, we monitored Per1:luciferase activity in unsynchronized HEK 293T cells transiently transfected with Clock/Bmal1. Per1:luc reporter activity is enhanced by Clock/Bmal1 transfection but repressed by the overexpression of either Cry1 or Chrono (Figure 3E). As has been previously demonstrated, CLOCKH360Y and BMAL1G612E missense mutants are resistant to CRY-mediated repression [49]. In contrast, CHRONO-mediated repression is unaffected by these point mutations (Figure 3E). The same pattern was observed in the expression of Nr1d1, an endogenous CLOCK/BMAL1 target (Figure S4). Alternatively, CHRONO knockdown augments Per1:luc reporter activity (Figure 3F).These data suggest that CHRONO and CRY1 have distinct binding sites and/or functional mechanisms.

Small interfering RNA (siRNA) mediated knockdown of C1orf51, the human homologue of Chrono, markedly dampened circadian oscillations in a genome-wide circadian screen [28]. Using NIH 3T3 cells expressing a Bmal1:dLuc reporter as a second model system, we tested the effects of four different short hairpin RNA (shRNA) constructs that reduced Chrono transcript expression and protein abundance (Figure S5). Comparing the pooled results to control demonstrates that Chrono knockdown reduces amplitude and increases circadian period (Figure 4A–F). To definitively establish the role of CHRONO in modulating circadian behavior, we obtained transgenic mice from the Knockout Mouse Project [50]. These mice incorporate a transgenic construct (Figure S6A) whereby the Chrono encoding region is flanked by Lox-P sites (Chrono^flx/flx) and utilizes a “knockout-first” cassette [51]. The transgenic allele is a knockout at the level of RNA processing. We mated heterozygous transgenic mice to obtain homozygous Chrono knockout mice (Chrono^flx/flx), wild-type littermate controls (Chrono^+/+), and heterozygotes (Chrono^flx/+). qPCR confirmed that, when compared to wild-type littermate controls, mRNA expression was halved in heterozygotes (Chrono^flx/+) and abolished to basal levels in homozygote knockouts (Chrono^flx/flx) (Figure S6B and C). As shown in Figure 4G and H, wild-type, heterozygous, and homozygous knockouts were all well entrained to the 12∶12 light∶dark (L∶D) cycle and maintained a 24-h period. Under free-running conditions, homozygous Chrono knockouts exhibited a statistically significant (p<0.05) ∼25-min increase in circadian period as compared to wild-type controls (Figure 4I). Heterozygous knockouts display an intermediate period. The magnitude of this period change is similar to that observed in Clock (∼20 min) [52], Per1 (∼40 min) [53], Per3 (∼30 min) [54], Nr1d1 (∼20 min) [6], Rorb (∼25 min) [55], and Npas2 (∼12 min) [56] knockout animals. These data strongly suggest that endogenous Chrono expression plays an important regulatory role in the mammalian circadian clock.

Light, however, does not appear to directly influence CHRONO expression in the SCN. The SCN microarray data of Jagannath et al. does not reveal a significant change in Chrono expression following a nocturnal light pulse [57]. Moreover, in our own experiments, the phase shifting response of Chrono knockout mice to light pulses at ZT16 or ZT22 are not significantly different from control. Thus the primary role of CHRONO in the circadian clock appears to be in modulating core oscillator function and output timing rather than oscillator entrainment.

In a recent report, BMAL2 was shown to function as a tissue-specific paralogue of BMAL1 [58]. However, CHRONO specifically binds BMAL1 and not BMAL2 (Figure 3A). Moreover, CHRONO functionally represses the transcriptional activity of the BMAL1/CLOCK complex but not the activity of the BMAL2/CLOCK complex (Figure 5A).

In order to identify the region of BMAL1 required for CHRONO binding, we generated mutant BMAL1 proteins with truncated N- or C-terminal regions (BMAL1^78–626, BMAL1^1–445) (Figure 5B) and tested their interaction with CHRONO using the mammalian two-hybrid assay. Deletion of the C-terminal domain of BMAL1 (BMAL1^1–445) completely abolished CHRONO binding, whereas deletion of the N-terminal domain (BMAL1^78–626) had no effect (Figure 5C). We next exploited the strong sequence homology between BMAL1 and BMAL2 to localize the CHRONO binding site within the BMAL1 C-terminal region. We swapped corresponding sections of the BMAL1 and BMAL2 C-terminal domains. As expected, the construct containing the N-terminal of BMAL1 and the full C-terminal of BMAL2 (BMAL1–BMAL2) did not interact with CHRONO in the two-hybrid assay and was relatively immune to CHRONO-mediated repression (Figures 5C–E). A chimeric protein including the N-terminal region of BMAL2 with the longer BMAL1 C-terminus (BMAL2–BMAL1#1) interacted with CHRONO and phenocopied wild-type BMAL1 with regard to CHRONO-mediated repression (Figure 5D–F). Sequence alignment between C-terminal domains of BMAL1 and BMAL2 reveals a region of poor alignment (514–594). Insertion of this unique region of the BMAL1 protein (514–594) into BMAL2 C-terminus rendered the chimeric protein (Bmal2–Bmal1#2) responsive to CHRONO-induced repression (Figure 5E–F). This CHRONO binding region is adjacent to, but distinct from, the CRY1 interacting terminus [59]. Thus, CHRONO functions as a specific transcriptional co-repressor of BMAL1 through interaction with a unique C-terminal domain adjacent to the CRY1 binding region. This domain is both necessary and sufficient for physical and functional interactions with CHRONO.

Previous studies suggested that CBP also binds to the BMAL1 C-terminus [59],[60]. Thus, we hypothesized that CHRONO might interfere with BMAL1–CBP binding. We generated plasmids encoding BMAL1 and CBP fused to the C- and N-terminal regions of the Venus YFP. We then utilized BiFC to visualize BMAL1–CBP interactions in HEK 293T cell nuclei. BMAL1–CBP complex formation induced a yellow BiFC signal (Figure 6A). Co-expression of native or S-tagged CHRONO severely dampened BMAL1–CBP complementation. Western blotting (Figure S7A) confirmed stable abundance of BMAL1 and CBP proteins, implicating altered binding as the source of the reduced BiFC signal. Lastly, the ability of CHRONO to interfere with BMAL1–CBP binding was verified by co-IP analysis showing that overexpression of intact CHRONO reduced BMAL1–CBP complex formation (Figure 6B).

A functional impairment in the ability of BMAL1 to recruit CBP is expected to reduce histone acetylation of CLOCK/BMAL1 target regions. To assess the influence of CHRONO on the histone acetyl-transferase activity of the BMAL1/CLOCK complex, we performed a ChIP study using an antibody targeting acetylated histone H3 lysine 9 (H3–K9) (Figure 6C). PCR was used to specifically evaluate H3–K9 acetylation near the Per1 promoter E-box. Control samples obtained from immortalized human osteosarcoma (U2OS) cells 24 and 36 h after dexamethasone synchronization demonstrated a temporal variation in target acetylation. U2OS cells overexpressing CHRONO demonstrated a blunted temporal profile in Per1 promoter H3–K9 acetylation, with loss of the increased acetylation normally observed 24 h after synchronization [61],[62].

In order to confirm that abrogated CBP/BMAL1 binding contributes to the CHRONO-mediated modulation of circadian dynamics, we constructed several CHRONO truncation mutants. All constructs that retained the 108–212 region reduced BMAL1–CBP binding as assessed by BiFC (Figure 6D and E). As has been previously demonstrated, overexpression of CBP, along with BMAL1 and CLOCK, enhances Per1:luc expression in unsynchronized cells (Figure 6F). Those same CHRONO constructs that abrogated BMAL1–CBP complex formation also repressed CBP-enhanced Per1:luc reporter activity (Figures 6E and F and S7B). This pattern of activity among CHRONO truncation mutants was further mirrored in their ability to colocalize with BMAL1 (Figure S7C). Stable expression of the constructs in synchronized cells reveals the same pattern in their ability to modulate circadian reporter expression (Figure S7D and E). Thus, the abrogation of the BMAL1–CBP binding provides a plausible mechanism whereby CHRONO might influence circadian dynamics.

In summary, our data demonstrate that Chrono (i) oscillates with a circadian frequency in multiple tissues, (ii) physically interacts with BMAL1 and PER2, (iii) specifically reduces BMAL1/CLOCK-mediated transcription independently of CRY1, (iv) affects the free-running circadian period of mice, and (v) interferes with BMAL1–CBP binding, functionally repressing the CLOCK/BMAL1 complex and modulating the circadian acetylation of target genes. Most importantly, CHRONO knockout mice display a long free-running circadian period similar to or more drastic than six other clock components. These data establish a role for Chrono in the mammalian circadian oscillator. Like CIPC [18], CHRONO appears unique to the vertebrate genome. Given the repressive function of CRY proteins, the evolutionary development of additional CLOCK/BMAL1 repressors in vertebrates highlights the importance of fine control of circadian rhythms. Transcriptional oscillations can differ in their amplitude, frequency, phase, basal expression, and waveform shape. The ability to independently control these characteristics likely requires multiple, tunable genetic parameters. The specificity of CHRONO-mediated repression for BMAL1 over BMAL2, along with tissue-specific variation in the expression of BMAL1 and BMAL2, may thus facilitate local tuning of circadian oscillations.

Of course, there remain important, unanswered questions with regard to the function of CHRONO in modulating circadian dynamics. Although the abrogation of BMAL1–CBP is a plausible mechanism for CHRONO-mediated repression, it may reflect only part of its circadian function. Moreover a nuanced understanding of how this repression leads to a period-lengthening phenotype in the knockout animal will likely require a greater understanding of kinetics and network compensation. Our BiFC data demonstrate that overexpressed CHRONO co-localizes with the CLOCK/BMAL1 complex in nuclear bodies. It was previously shown that BMAL1 recruits CBP primarily when localized to promyelocytic leukemia (PML) nuclear bodies [63]. Thus the interruption of the CBP–BMAL1 binding within these nuclear structures is consistent with the potent repression induced by CHRONO overexpression (Figure 3E). Indeed, while this work was in revision, Annayev et al. [64] also reported that CLOCK/BMAL1 transcription is efficiently repressed by CHRONO(GM129). Our experimental work adds both a description of the circadian locomotor phenotype of the Chrono knockout mouse and an understanding of the mechanism by which this repression is mediated.

Although our work focused on the interaction between CHRONO and BMAL1, CHRONO might also influence circadian physiology through its interaction with PER2. It was recently reported that PER2 also localizes to PML nuclear bodies [65]. The importance of CHRONO/PER2 binding (Figure 3A,C), both within this complex and more generally, remain unexplored. PER2 not only binds with cryptochromes but also interacts with nuclear receptors NR1D1 and RORA [66]. Our preliminary tests (Figure S8) show that overexpression of CHRONO enhances the PER2/NR1D1 complex formation. The recruitment of this established circadian repressor provides another mechanism for CHRONO-enhanced repression of the circadian network. The importance of CHRONO/PER2 binding and a broader analysis of the role of CHRONO in the circadian network will require further study.

The extent to which CLOCK can recruit CBP/P300 independently of BMAL1 also remains unclear [67]. Given the highly redundant structure of the circadian oscillator [68], the ability of CLOCK to recruit a co-activator hints that there may be a functional paralogue of CHRONO acting on the other half of the BMAL1/CLOCK complex. Perhaps most importantly, the knockout and targeted disruption of several other clock factors have been shown to not only influence circadian period but also downstream physiological changes in metabolism [69] and sleep homeostasis [70],[71]. More detailed phenotyping of CHRONO knockout mice will be required to identify any such deficits.

Machine learning has recently been applied to complex biological problems including drug discovery [72], protein translation [73], and gene interaction networks in yeast [74]. We used a simple form of probabilistic machine learning to integrate sparse existing data whose joint distribution is hypothesized to yield a more specific ranked list of candidate genes. Although follow-up experimentation is an important part of this process, the identification of Chrono reflects the ability of this approach to find genes regulating circadian behavior. To our knowledge, this is the first application of these methods to identify genes responsible for complex neurological behaviors. We anticipate that the investigation of other candidates will advance the understanding of circadian rhythms. Indeed, in addition to CHRONO, our initial screening of the top 25 novel candidates identified two other proteins that both bind clock components and modulate in vitro circadian oscillations. To facilitate the experimental characterization of these and other candidates, a more exhaustive candidate ranking is provided in Table S2. As bona fide clock components are discovered and high-quality datasets become available, exemplar distributions can be re-evaluated and feature metrics can be improved. Thus, this integrated computational and experimental approach presents a path for leveraging genome scale data to develop insight into circadian biology.

All animal experiments were performed with the approval of the Institutional Animal Care and Use Committee (IACUC Protocol Numbers 801906 and 803945).

Unless otherwise specified, all computations were done in the R programming environment [75].

The derivation of circadian evidence factors closely follows that for Bayes factors [32], and our strategy follows the Naïve Bayes Classifier approach of “learning” the feature distributions from the training data. We considered an individual feature described by metric , and a single arbitrary gene with observed metric value . The event space is divided in two disjoint events: and . These events correspond to a randomly selected gene having a metric value at least as extreme as or the randomly selected gene having a metric value less than . The events are labeled and , respectively. The use of an interval rather than a point allows us to regularize the sparse empirical data for the estimation. Each gene is assumed to belong to either the set of clock genes (Cgene) or the set of nonclock genes (NCgene).

By Bayes' Theorem:

and

Dividing (E1) by (E2) yields:

Substituting the definition of, the middle term of (E3) becomes:

The left-hand side of (E3) is the posterior odds of a gene being a core clock component conditional on observing a metric value greater than or equal to . The last term represents the general odds of clock gene membership without additional feature information. Thus, the posterior odds of a gene belonging to the set of clock genes (given a metric value greater than or equal to ) is equal to the product of and the a priori odds.

Our analysis included n = 5 clock gene features. For each metric , the event space is divided into two disjoint events— and for some —and these events are labeled and , respectively. Following the steps above:

The middle term in equation (E5) is the factor by which the a priori odds of clock gene membership must be adjusted to recover the posterior odds after all of the observed data. It represents the combined evidence factor () given all five features. Given the number of features, the training set of circadian clock components is too sparse to approximate the required joint distribution without some regularizing assumption. We follow the typical Naïve Bayes approach and show that, given conditional independence of the included features, is simply the product of the individual evidence factors.

By definition, random variableswith probability density functionsand joint probability density functionare conditionally independent given a random variableif and only if:

Using this definition and the definition of the events, the denominator ofcan be simplified:

Similarly, the numerator ofbecomes, and the cumulative evidence is equal to:

Given either the distribution of metric values among exemplar clock genes or the distribution among the genome at large, the probabilities of obtaining a metric greater than, or equal to, that observed was approximated with the ecdf() function in R. For any given gene and feature, the ratio of these probabilities was computed to obtain the value of. Metric values greater than the maximum value observed among exemplar clock components were assigned the same evidence factor as that maximum value. Combined evidence factors are the product of the feature-specific factors. If no data were available for a given gene and feature, this feature was ignored by setting the corresponding evidence factor to be 1. The ubiquity and homology metrics were both Boolean variables, and the standard Bayes factor formula was used for these features.

The use of combined evidence factors was compared with two prepackaged, supervised machine learning algorithms in the R programming environment: a Gaussian/Normal Naïve Bayes classifier within the “e1071” package [80] and a Flexible Naïve Bayes classifier [39] within the “klaR” package [81]. Probabilistic learning algorithms were preferred as they do not require a prior weighting of the importance of the various features [24]. For training, genes not in the exemplar clock group were labeled as nonclock genes, and the classifier was trained on the entire dataset. Genes were rank ordered on the posterior probability of clock gene membership after the model was applied to the data. For the Flexible Naïve Bayes implementation, kernel density estimation was performed with the default value for the “window parameter.” This default uses a heuristic formula to adjust the window of kernel density estimate based on the number of data points.

We sequentially removed all possible pairs of clock components from the exemplar distribution and retrained the various learning algorithms on the reduced exemplar sets, testing our ability to theoretically recover these known clock genes using different ranking cutoffs (). The three methods all had comparable performance using cutoffs less than ∼1,000, but the evidence factor method outperformed the other two beyond this point. The top candidates from all three methods show a very high degree of overlap (). We estimated the false discovery rate (FDR) of the Evidence Factor approach by combining the sensitivity analysis with an assumed total number of clock components to generate an expected number of true and false positives at different ranking thresholds (). Figure S2A Figure S2B Figure S2C

Preprocessed microarray data obtained from WT and Clock mutant animals as reported by Miller et al was downloaded from the Circa database [47] and replotted. A single apparent outlier from the SCN data (Mutant, original time point 46) is excluded from the plot as this value was greater than any other SCN expression value from WT or mutant animals, and ∼3× the replicate measure. Cel files from the Cry1/Cry2 double mutant were obtained from NIH GEO and normalized via GCRMA [76].

Exon-array cel files describing the transcriptional response of WT and melanopsin knockout animals to sham control and following a light pulse [57] were downloaded from NIH GEO. Data were extracted, annotated, quantile normalized, and log transformed at the gene level using the Affymetrix Expression Console package (v1.1). The probeset corresponding to Gm129 was then separately analyzed. In both WT and knockout animals, when compared to sham control, Chrono expression did not significantly change 30, 60, or 120 min after light pulse.

C- and N-terminal regions of an enhanced variant YFP called Venus were fused with identified constructs. Expression vectors of S-tagged full-length CHRONO (1–385), and its various deletion mutants were cotransfected with GFP–BMAL1 expression vector or BiFC fusion plasmids encoding VC–BMAL1, CLOCK–VN, or CBP–VN. At 16 h post-transfection, the cells were fixed with 4% paraformaldehyde in PBS and incubated with anti–S-tag (Bethyl Laboratories, Inc.) and anti-hC1orf51 (Santa Cruz Biotechnology) antibodies, followed by secondary antibodies conjugated to Alexa Fluor 568 (Invitrogen). Cells were visualized using fluorescein isothiocyanate and tetramethylrhodamine isothiocyanate filters in fluorescence microscopy.

A modified Aschoff type II procedure was used, facilitating the exposure of animals to light pulses before their free-running rhythms had drifted apart significantly [84],[85]. Animals were entrained to a 12∶12 L∶D cycle and then placed in constant darkness (D∶D) prior to a 30-min light pulse. The light pulses were initiated at zeitgeber times (ZTs) 16 or 22 on the second day of D∶D. Animals remained in DD for 7 d following the light pulse.

Daily activity onset times were determined using ClockLab Data Collection software (Actimetrics) and were exported for further analysis. The phase response was calculated as the difference between activity onset predictions as determined by prepulse and postpulse regression lines computed in R. The prepulse regression line was fit from activity onset data for 5 d prior to the light pulse. The postpulse regression line was determined from the first through seventh days in D∶D following the pulse [85].

We used 1 µg total RNA to generate cDNA with the High Capacity cDNA Archive Kit using the manufacturer's protocol (Applied Biosystems). qPCR reactions were performed using iTaq PCR mastermix (BioRad) in combination with gene expression assays (Applied Biosystems) on a 7800HT Taqman machine (Applied Biosystems). Importin 8 was used as an endogenous control for all experiments.

Mammalian two-hybrid constructs were generated by PCR with primers containing the flanking restriction sites that allow for in-frame cloning of the full-length ORF (not including the start ATG codon) into pACT or pBIND plasmids (Promega). The two-hybrid reporter plasmid pGL4P–4XUAS was generated by inserting 4× repeats of the Gal4 UAS binding sites into the pGL4P vector (Promega).

Epitope-tagged cDNAs were generated by PCR with primers containing the flanking restriction sites that allow for in-frame cloning of the full-length ORF (not including the start-ATG codon) into pFlag [49] or pTag3C plasmids (Stratagene).

For plasmids expressing S-tagged CHRONO (both wild-type and truncation mutants), full-length and truncated DNA fragments of the gene were amplified with upstream and downstream primers containing S-tag-encoding sequence (KETAAAKFERQHMDS) and were subcloned into pCMV Sport6 or pcDNA3.1 expression vectors (Invitrogen) using NotI and XhoI restriction enzymes.

The shRNA construct sequences were as follows: NS shRNA, CAACAAGATGAAGAGCACC; Sh234, GACTGGAGTTGCATCCTAT; Sh235, GAGCCAGCATTGGTGTCAT; Sh236, GACTTGGTTTCCTCACATA; Sh237, GGAGAACGTTATCTAGGAA; Sh238, GGAGCCTCGTTGCCACAGT; Sh239, GAACCTTGCTGCAGGTGGA; and Sh240, GTGTCATCCTTGTCCTCCA.

Sh 238, 239, and 240 were ultimately found to be ineffective by Western and/or PCR.

The Taqman probe identifiers were as follows: For Mus musculus: Arntl, Mm00500226_m1; Arntl2, Mm00549497_m1; Per1, Mm00501813_m1; Per2, Mm00478113_m1; Per3, Mm00478120_m1; Nr1d1, Mm00520708_m1; Chrono (Gm129), Mm01255906_g1; Importin 8, Mm01255158_m1. For Homo sapiens: Arntl, Hs00154147_m1; Arntl2, Hs 00368068_m1; Clock, Hs00231857_m1; Per1, Hs00242988_m1; Per2, Hs00256144_m1; Nr1d1, Hs00253876_m1; Chrono (C1orf51), Hs00328968_m1; Gapdh, Hs99999905_m1.